Polynomial solutions of differential-difference

arXiv:0902.0041v1 [[Link]] 31 Jan 2009

equations
Diego Dominici
Department of Mathematics
State University of New York at New Paltz
1 Hawk Dr. Suite 9
New Paltz, NY 12561-2443
USA
dominicd@[Link].
Kathy Driver
Department of Mathematics and Applied Mathematics
University of Cape Town
Private Bag X3, Rondebosch 7701
Cape Town
South Africa
[Link]@[Link].
Kerstin Jordaan
Department of Mathematics and Applied Mathematics
University of Pretoria
Pretoria, 0002
South Africa
kjordaan@[Link]
October 24, 2018

Abstract

1
 We investigate the zeros of polynomial solutions to the differential-
difference equation

Pn+1 (x) = An (x)Pn′ (x) + Bn (x)Pn (x), n = 0, 1, . . .

where An and Bn are polynomials of degree at most 2 and 1 respec-

tively. We address the question of when the zeros are real and simple
and whether the zeros of polynomials of adjacent degree are interlac-
ing. Our result holds for general classes of polynomials but includes se-
quences of classical orthogonal polynomials as well as Euler-Frobenius,
Bell and other polynomials.

AMS MOS Classification: 33C45, 42C05

Keywords: Interlacing properties; Zeros.

1 Introduction
Let {Pn (x)}∞
n=0 be the sequence of polynomials defined by

P0 (x) = 1
Pn+1 (x) = An (x)Pn′ (x) + Bn (x)Pn (x), n = 0, 1, . . . , (1)

where An (x) and Bn (x) are polynomials in x of degree at most 2 and 1

respectively. The best-known families of classical orthogonal polynomials
satisfy differential-difference equations of this type:

• Jacobi polynomials

(α,β) (2n + 2 + α + β) (x2 − 1) d (α,β)

Pn+1 (x) = P (x)
2(n + 1) (n + 1 + α + β) dx n
(2n + 2 + α + β) x + α − β (α,β)
+ Pn (x).
2(n + 1)

• Laguerre polynomials

(α) x d (α) α + n + 1 − x (α)

Ln+1 (x) = Ln (x) + Ln (x).
n + 1 dx n+1

2
 • Hermite polynomials
d
Hn+1 (x) = − Hn (x) + 2xHn (x).
dx
In these cases, it follows from the theory of orthogonal polynomials [9]
that the zeros of Pn (x) are real and simple and that the zeros of Pn+1 (x)
and Pn (x) are interlacing. This leads to the question of what can one say
about the zeros of a sequence of polynomials satisfying (1) which is not
orthogonal. Examples of such sequences include the Bell polynomials [1], the
Euler-Frobenius polynomials [5] and the so-called derivative polynomials [6].
A number of authors have investigated the properties of the zeros of
sequences of polynomials that are solutions of (1) but are not, in general,
orthogonal. In [10], Vertgeim considered polynomials generated by (1) with
An (x) = an x2 − bn , Bn (x) = αan x, α, an , bn > 0,
that generalize the Euler polynomials. In [3], [4], Dubeau and Savoie study
interlacing properties of the zeros of polynomial solutions of (1) with
An (x) = κn 1 − x2 , Bn (x) = −2κn rn x, rn > 0, κn 6= 0,

which contain the generalized Euler-Frobenius and the ultraspherical poly-

nomials as special cases. They also consider the Hermite-like polynomials
defined by (1) with
An (x) = κn , Bn (x) = −2κn x, κn 6= 0.
In [8], Liu and Wang analyze polynomial solutions of the equation
Pn+1 (x) = An (x)Pn′ (x) + Bn (x)Pn (x) + Cn (x)Pn−1 (x), n = 0, 1, . . . . (2)
By assuming that Pn (x) has strictly nonnegative coefficients (respectively
alternating in sign) and An (x) < 0 or Cn (x) < 0 for x ≤ 0 (respectively
x ≥ 0), , they show that {Pn (x)}∞ n=0 forms a Sturm sequence. Although
(2) is more general than (1), the conditions on the sign of the coefficients of
Pn (x) is a priori very difficult to check, except in rather isolated situations.
In this paper, we shall take an approach similar to that used in [10].
We establish criteria that ensure, for a sequence of polynomials {Pn (x)}∞n=0
satisfying (1), either that all zeros of Pn are real and simple, or the zeros
of Pn and Pn+1 are interlacing, or both, based on conditions that can be
checked directly from An and Bn . We present several interesting examples
and consider possible extensions.

3
2 Preliminary results
It is obvious that any sequence of polynomials is a solution of (1), if we allow
An (x) and Bn (x) to be (non unique) rational functions of x. The question
of characterizing which sequences of polynomials {Pn (x)}∞ n=0 are solutions of
(1), when An (x) and Bn (x) are polynomials in x, is addressed in the following
theorem.
Theorem 1 Suppose that Pn (x) is a polynomial with n simple zeros. Let
Pn+1 (xn,k )
Pn (xn,k ) = 0, yn,k = , k = 1, . . . , n. (3)
Pn′ (xn,k )
Then the following are equivalent:
i) There exist polynomials An (x) and Bn (x) of degree at most 2 and 1 re-
spectively such that {Pn (x)}∞
n=0 satisfies (1). For every n ≥ 3, An (x)
and Bn (x) are unique.
ii) For every n ≥ 2, there exists a quadratic polynomial An (x) such that
An (xn,k ) = yn,k , k = 1, . . . , n.

Proof. Evaluating (1) at the zeros of Pn (x), we clearly see that i) =⇒ ii).
Assume now that ii) holds. For n = 0, we have
P1 (x) = A0 (x)P0′ (x) + B0 (x)P0 (x) = B0 (x),
since P0 (x) = 1. Thus, A0 (x) can be any polynomial of degree at most 2 and
B0 (x) = P1 (x). For n = 1, let B1 (x) be an arbitrary polynomial of degree at
most 1 and define A1 (x) by
P2 (x) − B1 (x)P1 (x)
A1 (x) = .
P1′ (x)
Since P1′ (x) is a non-zero constant, we see that i) holds for n = 1. For n ≥ 2,
let An (x) be a quadratic polynomial such that
yn,k = An (xn,k ), k = 1, . . . , n.
Then Pn+1 (x) − An (x)Pn′ (x) is a polynomial of degree at most n + 1 that is
zero at each xn,k and therefore divisible by Pn (x). This yields the existence
of a Bn that is at most linear. When n ≥ 3, An is uniquely determined so
that i) holds.

4
Remark 2 The assumption that Pn (x) has n simple zeros is not very restric-
tive in the sense that the statement and proof of Theorem 1 can be modified
to cater for this possibility.

Example 3 Let Pn (x) be the family of orthonormal polynomials with respect

to the Freud-type weight
r
t2
 
2 1 4 2
w(x) =
exp −x + 2tx − ,
t K 1 t22 4
4

where Kν (z) is the Bessel function of the second kind. In this case, we have
[7]


1 x x2 + a2n+1 − t
An (x) = −
, Bn (x) =
,
4an+1 x2 + a2n+1 + a2n − t an+1 x2 + a2n+1 + a2n − t

where the numbers an are the coefficients in the three-term recurrence relation

xPn (x) = an+1 Pn+1 (x) + an Pn−1 (x)

satisfying the string equation

n = 4a2n a2n+1 + a2n + a2n−1 − t ,

n ≥ 0.

The initial values for an are

v
t2


K1 3
r u

tu 4 Γ 4
a0 = 0, a1 (t) = t 4
t2

− 1, a1 (0) = 1√ , (4)
2 K3 4 24 π
4

x
where a1 > 0 is chosen so that P1 (x) = a1
has unit norm.
For n = 5, we have
x5 − αx3 + βx
P5 (x) = ,
a1 a2 a3 a4 a5
with
α = a21 + a22 + a23 + a24 , β = a21 a23 + a21 a24 + a22 a24
and therefore, with the notation of (3),
p p p p
x5,1 = 0, x5,2 = ζ+ , x5,3 = − ζ+ , x5,4 = ζ− , x5,5 = − ζ−

5
 1 p 
ζ± = α ± α2 − 4β .
2
One can show, using a simple algebraic argument, that the polynomial inter-
polating the points (x5,k , y5,k ) , 1 ≤ k ≤ 5 has degree 4, unless a1 = 0, which
contradicts (4). We deduce from Theorem 1 that there are no polynomials
An (x)andBn (x) such that Pn (x) satisfies (1).

3 Main result
In our proofs, we will find it convenient to re-write the differential-difference
equation (1) in the form
An (x) d
Pn+1 (x) = [Kn (x)Pn (x)], (5)
Kn (x) dx
where Kn (x) is an ”integrating factor”, defined by
 x 
Bn (t) 
Z
Kn (x) = exp  dt . (6)
An (t)

Since An (x) and Bn (x) are polynomials, we can obtain all possible functions
Kn (x) by considering the location of the zeros of An (x) and Bn (x). This
leads to the following classification, where we define the extended real line
by (−∞, ∞) ∪ {−∞, ∞} and the notation f (±∞) = 0 means lim f (x) = 0.
x→±∞

Theorem 4 Let Kn (x) be defined by (6). Then, Kn (x) has at most 2 zeros
on the extended real line.
Proof. Since Kn (x) depends on the ratio Bn (x)/An (x), without loss of
generality, we can choose An (x) to be monic. We consider the possible cases
in turn.
1. Let An (x) = (x − λn ) (x − ξn ) , deg (B0 ) = 1 and Bn (x) = µn , µn 6= 0
for n ≥ 1. If λn 6= ξn , we have
  
µn x − λn
Kn (x) = exp ln .
λn − ξ n x − ξn
If ξn = λn , we have
 
µn
Kn (x) = exp − .
x − λn

6
 2. Let An (x) = (x − λn ) (x − ξn ) , and Bn (x) = κn (x − µn ) , κn 6= 0. If
ξn 6= λn 6= µn , we have
 
λn − µ n µn − ξ n
Kn (x) = exp κn ln (x − λn ) + κn ln (x − ξn ) .
λn − ξ n λn − ξ n

If ξn = λn 6= µn , we have
 
κnµ n − λn
Kn (x) = (x − λn ) exp κn .
x − λn

If ξn 6= λn = µn or ξn = λn = µn , we have

Kn (x) = (x − ξn )κn .

3. Let An (x) = x − λn , and Bn (x) = κn (x − µn ) , κn 6= 0. If λn 6= µn , we

have
Kn (x) = exp [κn x + κn (λn − µn ) ln (x − λn )] .
If λn = µn , we have
Kn (x) = exp (κn x) .

4. Let An (x) = 1, and Bn (x) = κn (x − µn ) , κn 6= 0. We have

h x i
Kn (x) = exp κn x − µn .
2

We are know ready to prove our main result.

Theorem 5 We denote the zeros of Pn (x) in increasing order by γi,n , i =

1, . . . , n. Let
−∞ ≤ αn+1 ≤ αn < βn ≤ βn+1 ≤ ∞
and Kn (x) be continuous on [αn , βn ] and differentiable on (αn , βn ) .Then,

(a) If Kn (x) = 0 only at x = αn and x = βn for 1 ≤ n ≤ N and α1 <

γ1,1 < β1 , then the zeros of Pn and Pn+1 interlace and are in the interval
(αn , βn ) for 1 ≤ n ≤ N.

7
 An (x)
(b) If Kn (x) = 0 only when x = αn (respectively βn ), K n (x)
= 0 when
x = βn (respectively αn ), βn < βn+1 for 1 ≤ n ≤ N (respectively
αn+1 < αn for 1 ≤ n ≤ N) and α1 < γ1,1 < β1 , then the zeros of Pn
and Pn+1 interlace and are in the interval (αn , βn ) for 1 ≤ n ≤ N.
An (x)
(c) If Kn (x) = 0 only when x = αn (respectively βn ), K n (x)
= 0 when
x = βn (respectively αn ) for 1 ≤ n ≤ N and α1 < γ1,1 ≤ β1 (respectively
α1 ≤ γ1,1 < β1 ), then all the zeros of Pn are real, simple and in the
interval (αn , βn ] (respectively [αn , βn ) ) for 1 ≤ n ≤ N.
An (x)
(d) If Kn (x) 6= 0, K n (x)
= 0 when x = αn , βn and αn+1 < αn < βn < βn+1
for 1 ≤ n ≤ N with α1 < γ1,1 < β1 , then the zeros of Pn and Pn+1
interlace for 1 ≤ n ≤ N.

Proof. We use the familiar extension of Rolle’s theorem to an infinite open

interval.
A1 (x) d
(a) When n = 1 we have P2 (x) = K 1 (x) dx
[K1 (x)P1 (x)] and K1 (x)P1 (x) = 0
at α1 , β1 and γ1,1 with α1 < γ1,1 < β. It follows from Rolle’s theorem
that P2 (x) = 0 at γ1,2 , γ2,2 with

α2 ≤ α1 < γ1,2 < γ1,1 < γ2,2 < β1 ≤ β2 .

Now, let 2 ≤ n ≤ N and assume that we have proved the result for
Pn (x). Since Kn (x)Pn (x) vanishes at αn , βn and γi,n for i = 1, . . . , n
with αn < γ1,n < . . . < γn,n < βn , Rolle’s theorem applied to (5) yields

αn+1 ≤ αn < γ1,n+1 < γ1,n < · · · < γn,n < γn+1,n+1 < βn ≤ βn+1

and the result follows.

An
(b) We prove the result for K n
(βn ) = 0 for 1 ≤ n ≤ N, the other case
A1 (x) d
being analogous. When n = 1 we have P2 (x) = K 1 (x) dx
[K1 (x)P1 (x)]
and K1 (x)P1 (x) = 0 at α1 and γ1,1 with α1 < γ1,1 < β1 and it follows
from Rolle’s theorem that α1 < γ1,2 < γ1,1 . The second zero of P2
coincides with β1 , so we have

α2 ≤ α1 < γ1,2 < γ1,1 < γ2,2 = β1 < β2 .

8
 Now, let 2 ≤ n ≤ N and assume that we have proved the result for
Pn (x). Since Kn (x)Pn (x) vanishes at αn and γi,n for i = 1, . . . , n with
αn < γ1,n < . . . < γn,n < βn , Rolle’s theorem applied to (5) yields

αn+1 ≤ αn < γ1,n+1 < γ1,n < · · · < γn,n .

Since the largest zero of Pn+1 coincides with βn we have

αn+1 ≤ αn < γ1,n+1 < γ1,n < · · · < γn,n < γn+1,n+1 = βn < βn+1

and the result follows.

An
(c) We prove the result for K n
(βn ) = 0 for 1 ≤ n ≤ N, the other case
A1 (x) d
being analogous. When n = 1 we have P2 (x) = K 1 (x) dx
[K1 (x)P1 (x)]
and K1 (x)P1 (x) = 0 at α1 and γ1,1 with α1 < γ1,1 ≤ β1 and it follows
from Rolle’s theorem that α1 < γ1,2 < γ1,1 . The second zero of P2
coincides with β1 , so we have

α2 ≤ α1 < γ1,2 < γ1,1 ≤ γ2,2 = β1 ≤ β2 .

Now, let 2 ≤ n ≤ N and assume that we have proved the result for
Pn (x). Since Kn (x)Pn (x) vanishes at αn and γi,n for i = 1, . . . , n with
αn < γ1,n < . . . < γn,n ≤ βn , Rolle’s theorem applied to (5) yields

αn < γ1,n+1 < γ1,n < · · · < γn−1,n < γn,n+1 < γn,n .

The largest zero of Pn+1 coincides with βn , so we have

αn+1 ≤ αn < γ1,n+1 < γ1,n < · · · < γn,n ≤ γn+1,n+1 = βn ≤ βn+1

and the result follows.

A1 (x) d
(d) When n = 1 we have P2 (x) = K 1 (x) dx
[K1 (x)P1 (x)] and P2 (x) = 0 at α1
and β1 with α1 = γ1,2 < γ2,2 = β1 . When n = 2, γ1,3 = β2 and γ3,3 = β2 .
A2 (x) d
Furthermore, P3 (x) = K 2 (x) dx
[K2 (x)P2 (x)] and Rolle’s theorem implies
that α1 = γ1,2 < γ2,2 = β1 . Hence

α2 = γ1,3 < α1 = γ1,2 < γ2,3 < γ2,2 = β1 < γ3,3 = β2 .

Now, let 3 ≤ n ≤ N and assume that we have proved the result

for Pn (x). The smallest and largest zero of Pn+1 is γ1,n+1 = αn and

9
 γn+1,n+1 = βn . The remaining n − 1 zeros are obtained by applying
Rolle’s theorem to the function inside the square brackets in (5)

αn = γ1,n+1 < γ1,n < γ2,n+1 < · · · < γn,n+1 < γn,n < γn+1,n+1 = βn .

4 Examples
We conclude by giving some examples where our results apply, highlighting
particular choices of An and Bn that give rise to known families of polyno-
mials.

Example 6 Let

An (x) = κn 1 − x2 ,


Bn (x) = −2κn rn x, rn > 0, κn 6= 0.

Then
rn
Kn (x) = x2 − 1
and Theorem 5 (a) applies. In [3], [4], the authors obtain the same result
using a different approach.

Example 7 The Bell polynomials Bn (x) are defined by

n
X
Bn (x) = Skn xk , n = 0, 1, . . . ,
k=0

where Skn is the Stirling number of the second kind. They satisfy the differential-
difference equation

Bn+1 (x) = x [B′n (x) + Bn (x)] ,

from which we obtain

An (x) = x, Kn (x) = ex .

In this case, we have αn = −∞ and βn = 0 and from Theorem 5 (c) it follows

that the zeros of Bn (x) are real and simple and lie in the interval (−∞, 0].

10
Example 8 Let Pn (x) be the family of polynomials defined by
 
−n, b
Pn (x) = (c)n 2 F1 x .
c

In this case, we have

An (x) = x(1 − x), Bn (x) = n + c − bx

and thus
Kn (x) = xn+c (x − 1)b−c−n .
Provided that n ∈ (−c, b − c), it follows from Theorem 5 (a) that the zeros
of Pn and Pn+1 are interlacing and lie in the interval (0, 1). The same result
is obtained in [2] using a different technique.

Acknowledgement 9 The work of D. Dominici was supported by a Hum-

boldt Research Fellowship for experienced researchers provided by the Alexan-
der von Humboldt Foundation. The work of K. Driver was supported by the
National Research Foundation of South Africa under grant number 2053730.
The work of K. Jordaan was supported by the National Research Foundation
of South Africa under grant number 2054423.

References
[1] L. Carlitz. Single variable Bell polynomials. Collect. Math., 14:13–25,
1962.

[2] K. Driver and K. Jordaan. Separation theorems for the zeros of certain
hypergeometric polynomials. J. Comput. Appl. Math., 199(1):48–55,
2007.

[3] F. Dubeau and J. Savoie. On interlacing properties of the roots of

orthogonal and Euler-Frobenius polynomials. In Approximation theory
VIII, Vol. 1 (College Station, TX, 1995), volume 6 of Ser. Approx.
Decompos., pages 185–191. World Sci. Publ., River Edge, NJ, 1995.

[4] F. Dubeau and J. Savoie. On the roots of orthogonal polynomials and

Euler-Frobenius polynomials. J. Math. Anal. Appl., 196(1):84–98, 1995.

11
 [5] F. G. Frobenius. Uber die Bernoullischen Zahlen und die Eulerischen
Polynome. Sitzungsber. Preuss. Akad. Wiss., 809–847, 1910.

[6] M. E. Hoffman. Derivative polynomials for tangent and secant. Amer.

Math. Monthly, 102(1):23–30, 1995.

[7] M. E. H. Ismail. Classical and quantum orthogonal polynomials in one

variable, volume 98 of Encyclopedia of Mathematics and its Applications.
Cambridge University Press, Cambridge, 2005.

[8] L. L. Liu and Y. Wang. A unified approach to polynomial sequences

with only real zeros. Adv. in Appl. Math., 38(4):542–560, 2007.

[9] G. Szegő. Orthogonal polynomials. American Mathematical Society,

Providence, R.I., fourth edition, 1975.

[10] B. A. Vertgeı̆m. Euler polynomials and their generalization. Sibirsk.

Mat. Zh., 33(2):164–167, 222, 1992.

12